Agentic AI is moving quickly from concept to deployment. Many organizations are already introducing agent-like capabilities into their environments, and this functionality is increasingly being built directly into enterprise software. CRM, ERP, service management, and industry platforms are beginning to include agents that automate workflows, coordinate tasks, and take action. Within those platforms, agentic AI is working.
However, when organizations deploy agentic AI across the enterprise, composed of multiple individual systems, they encounter obstacles that stop many implementations in their tracks.
Same pattern, more alarming risk
Analytics followed a similar path. Early value came from dashboards and reporting inside individual systems. But as organizations tried to analyze across the enterprise, limitations became clear. Data lived in silos, definitions did not align, and results had to be reconciled manually.
Agentic AI is now encountering the same structural challenge.
The difference is that analytics produced insight. Agentic AI produces action.
When data was inconsistent in the analytics era, it led to conflicting reports or delayed decisions. Someone had to step in, compare outputs, and determine what was correct. The issue was visible and often contained.
With agentic AI, the same inconsistency can trigger different actions across systems. An agent may update a customer record, trigger a transaction, or initiate a workflow based on data that is technically correct in one system but inconsistent with another. There is no natural pause for reconciliation.
Organizations that succeed with agentic AI will not be those that simply deploy more agents..
Similarly, whereas earlier AI systems – particularly copilots – operated within a workflow that required human review and approval before action was taken, creating a buffer in which inconsistencies can be resolved, agentic systems are designed to act autonomously. The focus shifts from assisting decisions to executing them.
This is where the risk changes.
Even the most advanced model will produce poor outcomes if it is operating on incomplete context, inconsistent definitions, or misaligned rules. In practice, what an agent does is shaped as much by the data it accesses and the constraints it operates under as by the model itself. Agentic AI is therefore not just a model problem or a data problem. It is a system problem.
The particular requirements of agentic AI
To operate reliably, agentic AI depends on three capabilities that are often incomplete today:
First, a contextualized view of the enterprise.
Agents need to understand how data connects across systems. For example, a ‘customer’ in one system may not map cleanly to an ‘account’ in another. Inventory levels may look sufficient in a warehouse system but not reflect commitments already made in orders or contracts. Individuals naturally reconcile these differences based on experience. Agents do not. Without a consistent view of how data relates across the organization, agents make decisions based on partial information.
Second, alignment to compliance and security policies.
Every action taken by an agent must follow rules — who can access data, what actions are allowed, and under what conditions. In many organizations, these rules are scattered across systems and enforced inconsistently. That is manageable when people are involved, because exceptions can be caught and corrected. When agents act autonomously, those inconsistencies can lead directly to actions that violate policy or create unintended exposure. Policies need to be applied consistently at the point where data is accessed and decisions are made.
Third, traceability.
When something goes wrong, organizations need to understand why. That means being able to trace what data was used, how it was interpreted, what rules were applied, and what action followed. In many environments today, this level of visibility is difficult to achieve across systems. Without it, diagnosing issues becomes slow and uncertain, and confidence in automated decisions declines.
The overarching challenge
Across all three of these areas, a common issue emerges: consistency of meaning.
In most organizations, definitions vary across systems. The same term can have slightly different meanings depending on where it is used. This has always been a challenge, but it was manageable when people were interpreting results.
Agents do not interpret meaning. They apply it.
Organizations that succeed with agentic AI will not be those that simply deploy more agents. They will be the ones that make it possible for those agents to operate with a consistent understanding of the enterprise, within clearly defined constraints, and with the ability to explain what they did and why.
Without that foundation, agentic AI does not just scale intelligence. It scales whatever inconsistencies already exist — more quickly and with less opportunity to intervene.
Be First to Comment